Towards Automated Image Quality Assessment in Ultrasound Imaging

Improving the image quality of ultrasound imaging has been a thriving research problem for many years. Many advanced beamforming algorithms including tissue adaptive and deep learning based frameworks have been proposed towards this. Standard image quality metrics such as Full Width Half Maximum (FWHM), Contrast Ratio (CR), Contrast to Noise Ratio (CNR), Generalized CNR (gCNR) and speckle statistics have been a few, which were frequently employed to gauge the performance of the algorithms. However, these quality metrics were employed on calibrated phantom images and were not extended to more realistic real-life ultrasound images. In this work, an approach towards fully automated image quality measurement for ultrasound imaging is presented. The approach consists of a lightweight neural network trained to detect the various regions of interest (e.g. FWHM region, CR region), which is followed by image processing to extract the relevant quality metrics such as the FWHM, CR, CNR, gCNR and speckle quality as per the established algorithms in the literature. The proposed object detection was able to detect the regions of interest with a mean average precision (mAP) of 0.53.